[liamsemeria]

Motivation

For long runs, the ability to resume a run from a results log was requested. Also the random forests had no filtering to avoid re-evaluating candidates, so it would end up using the same candidates as it converged.

Description

The resuming just starts from a new seed, determined by the number of trials that were executed.

I also changed the presets, since the filtering made the presets behave more similar to each other, and allowed a standard ucb acquisition converge similarly to the current random forests presets.
random-forest-default: used to be random-forest-explore but It behaves the best out of the 3 old presets
random-forest-ucb: the same random forest parameters as the previous default, but standard ucb
random-forest-ucb-explore: the same as above but more weight on std dev, could be better forruns with lots of trials
random: same as before

In terms of differences in performance there isn't a huge difference, but I would recommend using the ucb presets for runs with more than 400 trials.

Discussion

I kept the default to have the decay on the standard deviation in the ucb since it converges a little but faster, but we could change the default to be standard ucb instead.

[codecov]

Codecov Report

❌ Patch coverage is 89.51049% with 15 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/xtc/search/optimizers.py	88.40%	3 Missing and 5 partials ⚠️
src/xtc/search/explore.py	86.20%	2 Missing and 2 partials ⚠️
src/xtc/search/progress.py	50.00%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!

[liamsemeria]

@guillon updated it to restore from the last batch.